Yes, though you could condition it on a good human, or even yourself if you like.
The point is the human isn’t superintelligent. Hitler wouldn’t be able to hack a human brain, or invent an AGI. So an AI mimicking Hitler wouldn’t have those abilities either, and would have to solve the problem the hard way.
He was still just a man. He didn’t have super powers. He was a skilled orator and leader, with the right ideology, at the right time and place in history. But he wouldn’t be able to break out of an AI box. He wouldn’t be able to design a self replicating machine to destroy the world. A Hitler-in-a-box could be safely contained.
Enough skill in oratory (or something closely related) gets you unboxed. The question is how plausible it is that a superintelligent AI would have enough. (A related question is whether there’s such a thing as enough. There might not be, just as there’s no such thing as enough kinetic energy to let you escape from inside a black hole’s horizon, but the reported results of AI-Box games[1] suggest—though they certainly don’t prove—that there is.)
[1] The term “experiments” seems a little too highfalutin’.
[EDITED to add: I take it Houshalter is saying that Hitler’s known oratorical skills aren’t enough to convince him that H. would have won an AI-Box game, playing as the AI. I am inclined to agree. Hitler was very good at stirring up a crowd, but it’s not clear how that generalizes to persuading an intelligent and skeptical individual.]
Well for one, the human isn’t in a box trying to get out. So an AI mimicking a human isn’t going to say weird things like “let me out of this box!” This method is equivalent to writing Hitler a letter asking him a question, and him sending you an answer. That doesn’t seem dangerous at all.
Second, I really don’t believe Hitler could escape from a box. The AI box experiments suggest a human can do it, but the scenario is very different than a real AI box situation. E.g. no back and forth with the gatekeeper, and the gatekeeper doesn’t have to sit there for 2 hours and listen to the AI emotionally abuse him. If Hitler says something mean, the gatekeeper can just turn him off or walk away.
He could order such machine without leaving a box. He could grab power and start WW3 without leaving the box.
The ability to leave the box an make own copies is not necessary condition for dangerous AI.
If Hitler appeared today, even without being stuck inside a box, he would not be able to start WW3 or grab power. Hitler is not that powerful. And being stuck in a box would be a significant constraint on his power.
He was powerful enough to grab power in Germany in 13 years, even being inside the box of his cranial. If he will be a little bit better now (but not a superintelligence), just ideal psychopath, he (or his equivalent) could be even more effective.
Yes we can. Superintelligences have abilities that normal intelligences do not.
Imagine a game of chess. A good AI will make vastly different moves than a bad AI or a human. More skilled players would easily be detectable. They would make very different moves.
An AI that is indistinguishable from a human (to an even greater superintelligent AI) is not dangerous, because humans are not dangerous. Just like a chess master that is indistinguishable from a regular player wouldn’t win many games.
Can’t remember ad hoc; but if superintelligence is able to do anything, it could easily pretend to be more stupid than it is. May be only “super super intelligence” could solve him.
But it also may depend of the length of the conversation. If it say just Yes or No once, we can’t decide, if it say longer sequences we could conclude something, but for any length of sentences is maximum level of intelligence that could be concluded from it.
The opportunities for detecting superintelligence would definitely be rarer if the superintelligence is actively trying to conceal the status.
What about in the case where there is no attempted concealment? Or even weaker, where the AI voluntary submits to arbitrary tests. What tests would we use?
Presumably we would have a successful model of human intelligence by that point. It’s interesting to think about what dimensions of intelligence to measure. Number of variables simultaneously optimized? Optimization speed? Ability to apply nonlinear relationships? Search speed in a high dimensional, nonlinear solution space? I guess it is more the ability to generate appropriate search spaces in the first place. Something much simpler?
limited proxies—yes, well said.
also I would add solving problems which humans were unable to solve for long: aging, cancer, star travel, word peace, resurrection of dead.
Or we could ask these AI to create the scale. We could use also its size to estimate power, like number on neurons.
But real test needs to be powerful as well as universal optimization problem, something like ability to crack complex encryption or Go game.
I created a list of steps or milestones of the future AI and we could use similar list to estimate level of current super AI.
AI autopilot. Tesla has it already.
AI home robot. All prerequisites are available to build it by 2020 maximum. This robot will be able to understand and fulfill an order like ‘Bring my slippers from the other room’. On its basis, something like “mind-brick” may be created, which is a universal robot brain able to navigate in natural space and recognize speech. Then, this mind-brick can be used to create more sophisticated systems.
AI intellectual assistant. Searching through personal documentation, possibility to ask questions in a natural language and receive wise answers. 2020-2030.
AI human model. Very vague as yet. Could be realized by means of a robot brain adaptation. Will be able to simulate 99% of usual human behavior, probably, except for solving problems of consciousness, complicated creative tasks, and generating innovations. 2030.
AI as powerful as an entire research institution and able to create scientific knowledge and get self-upgraded. Can be made of numerous human models. 100 simulated people, each working 100 times faster than a human being, will be probably able to create AI capable to get self-improved faster, than humans in other laboratories can do it. 2030-2100
5a Self-improving threshold. AI becomes able to self-improve independently and quicker than all humanity
5b Consciousness and qualia threshold. AI is able not only pass Turing test in all cases, but has experiences and has understanding why and what it is.
Mankind-level AI. AI possessing intelligence comparable to that of the whole mankind. 2040-2100
AI with the intelligence 10 – 100 times bigger than that of the whole mankind. It will be able to solve problems of aging, cancer, solar system exploration, nanorobots building, and radical improvement of life of all people. 2050-2100
Jupiter brain – huge AI using the entire planet’s mass for calculations. It can reconstruct dead people, create complex simulations of the past, and dispatch von Neumann probes. 2100-3000
Galactic kardashov level 3 AI. Several million years from now.
But human still could be very bad, like Hitler.
Or Donald Trump.
(It’s just a joke...or is it?)
Yes, though you could condition it on a good human, or even yourself if you like.
The point is the human isn’t superintelligent. Hitler wouldn’t be able to hack a human brain, or invent an AGI. So an AI mimicking Hitler wouldn’t have those abilities either, and would have to solve the problem the hard way.
But with Hitler’s mass hysteria skills, the AI could lead others to very crazy ends.
Yes, you do not need to be superintelligent to kill everybody. The later is simplier task.
He was still just a man. He didn’t have super powers. He was a skilled orator and leader, with the right ideology, at the right time and place in history. But he wouldn’t be able to break out of an AI box. He wouldn’t be able to design a self replicating machine to destroy the world. A Hitler-in-a-box could be safely contained.
Isn’t the skill of oratory precisely the skill that gets you unboxed?
Enough skill in oratory (or something closely related) gets you unboxed. The question is how plausible it is that a superintelligent AI would have enough. (A related question is whether there’s such a thing as enough. There might not be, just as there’s no such thing as enough kinetic energy to let you escape from inside a black hole’s horizon, but the reported results of AI-Box games[1] suggest—though they certainly don’t prove—that there is.)
[1] The term “experiments” seems a little too highfalutin’.
[EDITED to add: I take it Houshalter is saying that Hitler’s known oratorical skills aren’t enough to convince him that H. would have won an AI-Box game, playing as the AI. I am inclined to agree. Hitler was very good at stirring up a crowd, but it’s not clear how that generalizes to persuading an intelligent and skeptical individual.]
Well for one, the human isn’t in a box trying to get out. So an AI mimicking a human isn’t going to say weird things like “let me out of this box!” This method is equivalent to writing Hitler a letter asking him a question, and him sending you an answer. That doesn’t seem dangerous at all.
Second, I really don’t believe Hitler could escape from a box. The AI box experiments suggest a human can do it, but the scenario is very different than a real AI box situation. E.g. no back and forth with the gatekeeper, and the gatekeeper doesn’t have to sit there for 2 hours and listen to the AI emotionally abuse him. If Hitler says something mean, the gatekeeper can just turn him off or walk away.
He could order such machine without leaving a box. He could grab power and start WW3 without leaving the box. The ability to leave the box an make own copies is not necessary condition for dangerous AI.
If Hitler appeared today, even without being stuck inside a box, he would not be able to start WW3 or grab power. Hitler is not that powerful. And being stuck in a box would be a significant constraint on his power.
He was powerful enough to grab power in Germany in 13 years, even being inside the box of his cranial. If he will be a little bit better now (but not a superintelligence), just ideal psychopath, he (or his equivalent) could be even more effective.
We can’t judge based on behaviour that some one is superintelligent or not.
Yes we can. Superintelligences have abilities that normal intelligences do not.
Imagine a game of chess. A good AI will make vastly different moves than a bad AI or a human. More skilled players would easily be detectable. They would make very different moves.
But in some games it is better to look more stupid in the begging. Like poker, espionage and AI box experiment.
An AI that is indistinguishable from a human (to an even greater superintelligent AI) is not dangerous, because humans are not dangerous. Just like a chess master that is indistinguishable from a regular player wouldn’t win many games.
It may be indistinguishable until it gets our of the building. Recent movie Ex Machine had such plot.
The AI doesn’t want to escape from the building. It’s utility function is basically to mimic humans. It’s a terminal value, not a subgoal.
But most humans would like to escape from any confinement
I wonder if this is true in general. Have you read a good discussion on detecting superintelligence?
Can’t remember ad hoc; but if superintelligence is able to do anything, it could easily pretend to be more stupid than it is. May be only “super super intelligence” could solve him. But it also may depend of the length of the conversation. If it say just Yes or No once, we can’t decide, if it say longer sequences we could conclude something, but for any length of sentences is maximum level of intelligence that could be concluded from it.
The opportunities for detecting superintelligence would definitely be rarer if the superintelligence is actively trying to conceal the status.
What about in the case where there is no attempted concealment? Or even weaker, where the AI voluntary submits to arbitrary tests. What tests would we use?
Presumably we would have a successful model of human intelligence by that point. It’s interesting to think about what dimensions of intelligence to measure. Number of variables simultaneously optimized? Optimization speed? Ability to apply nonlinear relationships? Search speed in a high dimensional, nonlinear solution space? I guess it is more the ability to generate appropriate search spaces in the first place. Something much simpler?
Probably winning humans in ALL known domains, including philosophy, poetry, love, power.
Although we use limited proxies (e.g., IQ test questions) to estimate human intelligence.
limited proxies—yes, well said. also I would add solving problems which humans were unable to solve for long: aging, cancer, star travel, word peace, resurrection of dead.
I mean, the ability to estimate the abilities of superintelligences appears to be an aspect of reliable Vingean reflection.
Or we could ask these AI to create the scale. We could use also its size to estimate power, like number on neurons. But real test needs to be powerful as well as universal optimization problem, something like ability to crack complex encryption or Go game.
I created a list of steps or milestones of the future AI and we could use similar list to estimate level of current super AI.
AI autopilot. Tesla has it already.
AI home robot. All prerequisites are available to build it by 2020 maximum. This robot will be able to understand and fulfill an order like ‘Bring my slippers from the other room’. On its basis, something like “mind-brick” may be created, which is a universal robot brain able to navigate in natural space and recognize speech. Then, this mind-brick can be used to create more sophisticated systems.
AI intellectual assistant. Searching through personal documentation, possibility to ask questions in a natural language and receive wise answers. 2020-2030.
AI human model. Very vague as yet. Could be realized by means of a robot brain adaptation. Will be able to simulate 99% of usual human behavior, probably, except for solving problems of consciousness, complicated creative tasks, and generating innovations. 2030.
AI as powerful as an entire research institution and able to create scientific knowledge and get self-upgraded. Can be made of numerous human models. 100 simulated people, each working 100 times faster than a human being, will be probably able to create AI capable to get self-improved faster, than humans in other laboratories can do it. 2030-2100
5a Self-improving threshold. AI becomes able to self-improve independently and quicker than all humanity 5b Consciousness and qualia threshold. AI is able not only pass Turing test in all cases, but has experiences and has understanding why and what it is.
Mankind-level AI. AI possessing intelligence comparable to that of the whole mankind. 2040-2100
AI with the intelligence 10 – 100 times bigger than that of the whole mankind. It will be able to solve problems of aging, cancer, solar system exploration, nanorobots building, and radical improvement of life of all people. 2050-2100
Jupiter brain – huge AI using the entire planet’s mass for calculations. It can reconstruct dead people, create complex simulations of the past, and dispatch von Neumann probes. 2100-3000
Galactic kardashov level 3 AI. Several million years from now.
All-Universe AI. Several billion years from now